When humans say something is "probable" or "likely," we usually share a general sense of what that means, even if it’s not precise. But when an AI chatbot like ChatGPT uses those same words, it doesn’t interpret them the way people do. A recent study published in npj Complexity reveals that large language models often misalign with human understanding when expressing uncertainty. The research focused on words that convey estimative probability—terms like "maybe," "probably," and "almost certain"—and compared how humans and AI models assign numerical values to them. The findings show that while AI models generally agree with humans on extreme cases—such as "impossible" or "certain"—they diverge significantly on more ambiguous terms. For example, an AI might associate "likely" with an 80% chance, while most people interpret it as closer to 65%. This mismatch arises because humans rely on context, experience, and social cues to understand uncertainty, whereas AI models are trained on vast amounts of text where the same word can be used inconsistently. As a result, the model may average out these conflicting uses, leading to estimates that don’t match human intuition. The study also uncovered that AI models are influenced by subtle changes in language. When prompts shifted from "he" to "she," the AI’s probability estimates often became more rigid, reflecting gender biases present in the training data. Similarly, switching from English to Chinese led to shifts in probability judgments, likely due to cultural and linguistic differences in how uncertainty is expressed. This misalignment is more than just a semantic quirk—it’s a serious concern for safety and trust. As AI systems are increasingly used in critical areas like healthcare, policy-making, and scientific communication, how they convey risk can have real-world consequences. If an AI tells a doctor a side effect is "unlikely" but means something closer to 40% while the doctor interprets it as less than 10%, the decision-making process could be dangerously flawed. Researchers have long studied how people quantify uncertainty, beginning with intelligence analysts in the 1960s. More recently, efforts have focused on understanding the internal workings of AI models. This study takes a new approach by treating human-AI interaction as a system where meaning can degrade over time, going beyond simple measures of intelligence to assess alignment. Some researchers are testing techniques like chain-of-thought prompting—asking AI to explain its reasoning—to improve accuracy. However, the study found that even with detailed reasoning, models often fail to close the gap between statistical likelihood and verbal labels. Moving forward, the goal is to develop AI that not only predicts the next word but truly understands the weight of uncertainty it expresses. Creating consistent, reliable metrics—so that a 10% chance always maps to the same word—will be essential. As AI becomes more integrated into daily life, ensuring that "probably" means "probably" is not just a linguistic issue, but a foundational step toward building trustworthy, human-centered AI systems.